openfold.utils.chunk_utils

Classes

ChunkSizeTuner([max_chunk_size])

Functions

chunk_layer(layer, inputs, chunk_size, ...)

Implements the "chunking" procedure described in section 1.11.8.
class ChunkSizeTuner(max_chunk_size=512)
tune_chunk_size(representative_fn, args, min_chunk_size)
Parameters:
Return type:

int

chunk_layer(layer, inputs, chunk_size, no_batch_dims, low_mem=False, _out=None, _add_into_out=False)

Implements the “chunking” procedure described in section 1.11.8.

Layer outputs and inputs are assumed to be simple “pytrees,” consisting only of (arbitrarily nested) lists, tuples, and dicts with torch.Tensor leaves.

Parameters:

  • layer (Callable) – The layer to be applied chunk-wise

  • inputs (Dict[str, Any]) – A (non-nested) dictionary of keyworded inputs. All leaves must be tensors and must share the same batch dimensions.

  • chunk_size (int) – The number of sub-batches per chunk. If multiple batch dimensions are specified, a “sub-batch” is defined as a single indexing of all batch dimensions simultaneously (s.t. the number of sub-batches is the product of the batch dimensions).

  • no_batch_dims (int) – How many of the initial dimensions of each input tensor can be considered batch dimensions.

  • low_mem (bool) – Avoids flattening potentially large input tensors. Unnecessary in most cases, and is ever so slightly slower than the default setting.

  • _out (Any | None)

  • _add_into_out (bool)

Returns:

The reassembled output of the layer on the inputs.

Return type:

Any